Integrating computer vision and machine learning technologies for model building to quantify intermuscular fat content in salmonid fillets

Abstract

The object of present study was to build a rapid and efficient method to quantify the intermuscular fat (IMF) content in salmonid fillets, which directly determines the fillet quality. Totally, 204 images of rainbow trout fillets were acquired for the preliminary IMF estimation through traditional RGB distribution analysis. However, its performance was unsatisfactory (R² = 0.61). Therefore, 9 color features and 7 composite features were further extracted from images to train linear (SR, EN) and nonlinear (RF, DNN) models based on computer vision and machine learning technologies. All models achieved accuracy>63 % and R² > 0.85. The RF model was considered the most robust with R² = 0.91 and accuracy = 79 %. Furthermore, the robust RF model was applied to quantify IMF of 120 additional fillets. IMF content showed significant association with body sizes, sex, and genetic backgrounds. It provided the first robust method for quantifying IMF content in salmonid fillets, with advantages of efficiency, accuracy, practicality, and reliability.